The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process untrusted external data, which is a common vector for indirect prompt injection attacks.
Ingestion points: In SKILL.md and references/company-annual.md, the agent is directed to read bank exports (CSV), Stripe transaction records, and vendor invoices (PDF/Image).
Boundary markers: The instructions mandate the creation of 'evidence tables' and the use of placeholders like '[FEHLT]' for missing data, which provide structural boundaries but are not cryptographic or parser-level protections.
Capability inventory: The agent can read files, perform calculations, and prepare draft letters for the human user. It does not have autonomous network write capabilities or direct tool execution for filing.
Sanitization: The skill does not provide specific instructions for sanitizing or escaping the text content of ingested financial documents before processing.
[COMMAND_EXECUTION]: The skill includes instructions for running specific shell commands related to its primary purpose.
Evidence: In references/elster-access.md, the skill provides an openssl command to check certificate expiry: 'openssl pkcs12 -in cert.pfx -nokeys -passin env:PFX_PW -legacy'.
Evidence: It mentions the use of 'myebilanz' (a tool for E-Bilanz filing) which may be run via a shell or Wine environment.
[EXTERNAL_DOWNLOADS]: The skill references several official and well-known government domains for tax filing and legal verification.
Evidence: The skill links to and fetches information from official portals including elster.de, bundesfinanzministerium.de, gesetze-im-internet.de, and formulare-bfinv.de.
Assessment: These references target official German government and finance services and are documented neutrally.

german-taxes-elster